VLA-Retain: Benchmarking Cross-Dimension Effects of Vision Tuning in VLAs

Borui Xu; Han Chen; Bingsheng He; Ming Wu; Jason Zeng; Michael Heinrich

cs.CVOct 20, 2026

VLA-Retain: Benchmarking Cross-Dimension Effects of Vision Tuning in VLAs

Authors: Borui Xu, Han Chen, Bingsheng He, Ming Wu, Jason Zeng, Michael Heinrich

Organizations: National University of Singapore

Abstract

Vision-language-action (VLA) models transfer knowledge from large-scale pretraining into robot control, but their manipulation performance remains sensitive to changes in observation conditions, which recent work addresses by fine-tuning the visual module on target-domain data. However, it remains unclear how such adaptation affects robustness beyond the target condition, as existing robustness benchmarks only evaluate models without fine-tuning and adaptation studies only report the target gain. Measuring this cross-dimension effect requires separating target gains from off-target changes, isolating adaptation effects from architectural differences, ruling out generic forgetting, and comparing tuning choices. Therefore, we introduce VLA-Retain, a benchmark that compares five open VLA models, tuned on Camera-view or Sensor-noise data, with their own base checkpoints on LIBERO-Plus dataset spanning seven perturbation dimensions. Our study reveals four findings. First, targeted vision tuning is not local, and its target gain does not predict retention. Mean off-target gain is negative in 8 of 10 model--domain configurations. Under Camera-view tuning, X-VLA gains $69.4$ percentage points on the target but loses $4.6$ points on average elsewhere, whereas OpenVLA-OFT gains $39.3$ and $1.4$ points, respectively. Second, these regressions are not explained only by performance loss on the same tasks without perturbations. Third, broader training coverage raises overall gain over target-specific training but does not eliminate dimension-level regressions. Fourth, tuning strategies are model-dependent, as DELTA-style feature preservation raises X-VLA's mean off-target gain to $+1.2$ points but lowers OpenVLA-OFT's to $-23.1$ . Overall, these findings show that target-domain success alone is insufficient and highlight cross-dimensional retention as a key criterion for targeted VLA adaptation.

Figures & tables

Figure 1: Overview of VLA-Retain. Target-domain data update only the visual module, but the updated module processes observations under every deployment condition. VLA-Retain therefore evaluates each base–tuned pair on the same episodes across seven perturbation dimensions and reports both target improvement and off-target retention.

Model	Visual backbone	Action formulation	Visual share
OpenVLA-OFT ( Kim et al., 2025a )	SigLIP + DINOv2	Regression head	$10.64\%$
$\pi_{0.5}$ ( Black et al., 2025 )	SigLIP	Flow matching	$11.40\%$
InternVLA-A1.5 ( Ma et al., 2026 )	Qwen3.5 ViT	Flow matching	$12.34\%$
NORA-1.5 ( Hung et al., 2025 )	Qwen2.5-VL ViT	Flow matching + FAST	$16.77\%$
X-VLA ( Zheng et al., 2026 )	DaViT	Soft-prompted flow matching	$41.26\%$

Table 1: Architectural coverage of the core benchmark. Visual share denotes the fraction of total parameters contained in the visual backbone and connector.

Dimension	Spatial	Object	Goal	LIBERO-10	Total
Camera view	376	396	408	419	1,599
Robot state	350	398	409	393	1,550
Language	390	354	410	383	1,537
Lighting	292	297	279	274	1,142
Background	258	248	281	289	1,076
Sensor noise	351	422	379	449	1,601

Table 2: Evaluation composition across perturbation dimensions and LIBERO task suites.

Camera-view tuning
	Target success			Off-target change		Overall success
Model	Base	Tuned	$\bm{G^{\mathrm{target}}}$	$\bm{G^{\mathrm{off}}}$	$\bm{G^{\mathrm{worst}}}$	Base	Tuned	$\bm{G^{\mathrm{all}}}$
OpenVLA-OFT	$55.2$	$94.5$	$+39.3$	$+1.4$	$-2.8$	$68.3$	$75.8$	$+7.5$
$\pi_{0.5}$	$70.7$	$92.7$	$+22.0$	$-0.3$	$-3.9$	$84.0$	$87.0$	$+3.0$
InternVLA-A1.5	$80.8$	$83.4$	$+2.6$	$-5.2$	$-12.4$	$83.7$	$79.5$	$-4.2$
NORA-1.5	$6.0$	$34.7$	$+28.7$	$-10.1$	$-17.8$	$47.4$	$43.8$	$-3.7$

Table 3: Target, off-target, and overall results under targeted vision tuning. Success rates are reported as percentages and signed gains are reported in percentage points.

Figure 2: Cross-dimensional success rate changes after targeted vision tuning. Each cell reports a percentage-point change. Blue denotes improvement, red denotes regression, and the outlined column marks the target perturbation.

Figure 3: Statistical uncertainty in off-target retention. Points show mean off-target gains, and horizontal lines show pointwise

95\%

confidence intervals from a paired cluster bootstrap with 20,000 resamples over the 40 source tasks. Red intervals lie below zero, blue intervals lie above zero, and gray intervals overlap zero.

Figure 4: Source-adjusted off-target changes. Results for (a) Camera-view and (b) Sensor-noise tuning. Each cell subtracts the corresponding change on original LIBERO; negative values indicate additional perturbation-specific regression.

Figure 5: Effect of training-data coverage. (a) Overall gains under Camera-view, Sensor-noise, and Mixed training. (b) Dimension-level and overall changes after Mixed training.

Figure 6: Target–retention trade-offs of tuning strategies. (a–b) Tuning scope. (c–d) WiSE-FT.

Model	Strategy	$\bm{G^{\mathrm{target}}}$	$\bm{G^{\mathrm{off}}}$	$\bm{G^{\mathrm{all}}}$
X-VLA	Dual LoRA	$+69.4$	$-4.6$	$+7.0$
X-VLA	DELTA-style	$+62.7$	$+1.2$	$+10.9$
OpenVLA-OFT	Dual LoRA	$+39.3$	$+1.4$	$+7.5$
OpenVLA-OFT	DELTA-style	$+13.6$	$-23.1$	$-16.6$

Table 4: Camera-view results for Dual LoRA and DELTA-style feature preservation.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Category	Configuration	Operation
Reference	Dual LoRA	Rank-32 LoRA on visual backbone and connector
Tuning scope	Backbone-only LoRA	Rank-32 LoRA on visual backbone; connector frozen
Tuning scope	Connector-only LoRA	Rank-32 LoRA on connector; visual backbone frozen
Retention	WiSE-FT	Base–Dual-LoRA visual-weight interpolation
Retention	DELTA-style	Dual LoRA with post-connector feature preservation

Appendix

Table 5: Visual tuning configurations used in the strategy comparison.

Figure 7: Architectures and visual interfaces of the five VLA models. The reference configuration updates the visual backbone and connector and freezes the multimodal backbone and action module.

Model	Vision params.	Total params.	Vision share
OpenVLA-OFT	802.3M	7.54B	10.64%
$\pi_{0.5}$	412.4M	3.62B	11.40%
InternVLA-A1.5	331.4M	2.69B	12.34%
NORA-1.5	668.7M	3.99B	16.77%
X-VLA	362.8M	0.88B	41.26%

Appendix

Table 6: Model parameter scales. Vision parameters include the visual backbone and connector.

	Base	Camera-view tuned		Sensor-noise tuned
Model	Success (%)	Success (%)	Change (pp)	Success (%)	Change (pp)
OpenVLA-OFT	$97.20$	$96.65$	$-0.55$	$96.60$	$-0.60$
$\pi_{0.5}$	$97.35$	$96.65$	$-0.70$	$96.05$	$-1.30$
InternVLA-A1.5	$98.25$	$97.95$	$-0.30$	$97.30$	$-0.95$
NORA-1.5	$94.70$	$81.30$	$-13.40$	$92.75$	$-1.95$
X-VLA	$97.50$	$93.70$	$-3.80$	$96.45$	$-1.05$

Appendix

Table 8: Original LIBERO success before and after targeted vision tuning. Success rates are percentages and changes are percentage points.

Model	Frozen teacher	Matched representation
X-VLA	Base visual path and image projector	image_proj_norm output
OpenVLA-OFT	Base vision backbone and projector	Projector output

Appendix

Table 9: Model-specific feature taps for DELTA-style preservation.

OpenVLA-OFT
Perturbation	Spatial	Object	Goal	LIBERO-10	All suites
Camera view	$48.7/98.9$	$67.9/99.7$	$59.3/93.6$	$45.1/86.4$	$55.2/94.5$
Robot state	$23.1/24.6$	$19.8/19.1$	$21.3/21.8$	$36.6/36.9$	$25.2/25.5$
Language	$95.9/96.4$	$99.7/98.6$	$58.0/59.0$	$95.3/95.0$	$86.5/86.6$
Lighting	$97.9/98.6$	$99.3/100.0$	$97.1/93.9$	$85.0/90.9$	$95.0/96.0$
Background	$88.0/88.4$	$87.5/81.0$	$84.0/88.6$	$83.0/91.0$	$85.5/87.5$

Appendix

Table 10: Camera-view tuning results by task suite. Each cell is Base/Tuned (%).

OpenVLA-OFT
Perturbation	Spatial	Object	Goal	LIBERO-10	All suites
Camera view	$48.7/53.2$	$67.9/73.2$	$59.3/55.6$	$45.1/39.6$	$55.2/55.2$
Robot state	$23.1/20.0$	$19.8/21.6$	$21.3/21.0$	$36.6/37.7$	$25.2/25.2$
Language	$95.9/93.8$	$99.7/99.4$	$58.0/59.3$	$95.3/95.6$	$86.5/86.3$
Lighting	$97.9/96.9$	$99.3/100.0$	$97.1/89.2$	$85.0/90.9$	$95.0/94.4$
Background	$88.0/89.1$	$87.5/90.3$	$84.0/84.3$	$83.0/85.5$	$85.5/87.2$

Appendix

Table 11: Sensor-noise tuning results by task suite. Each cell is Base/Tuned (%).

X-VLA
Strategy	Cam.	Robot	Lang.	Light	Back.	Noise	Layout	Overall
Connector only	$+36.5$	$-4.1$	$-13.3$	$-6.0$	$-5.2$	$-9.6$	$-4.0$	$-0.2$
Backbone only	$+69.4$	$-0.1$	$-4.9$	$-1.8$	$-2.2$	$-10.6$	$-3.7$	$+7.6$
Dual LoRA	$+69.4$	$+0.5$	$-6.6$	$-2.2$	$-3.9$	$-10.7$	$-4.9$	$+7.0$
WiSE-FT $\alpha=0.25$	$+11.6$	$-0.1$	$+0.3$	$+1.7$	$+0.9$	$+0.7$	$+2.0$	$+2.6$
WiSE-FT $\alpha=0.50$	$+32.0$	$+1.4$	$+1.5$	$+0.8$	$+2.0$	$-0.6$	$+1.7$	$+6.0$

Appendix

Table 12: Dimension-level gains under Camera-view tuning and retention strategies. All values are percentage points.

Figure 8: Camera-view checkpoint dynamics. Target, mean off-target, and overall gains are measured relative to the base model on a fixed balanced subset. All gains are reported in percentage points. The target and off-target changes do not evolve monotonically together.

Model	Seed	$\bm{G^{\mathrm{target}}}$	$\bm{G^{\mathrm{off}}}$	$\bm{G^{\mathrm{all}}}$
X-VLA	42	$+70.0$	$-4.0$	$+6.6$
X-VLA	43	$+73.8$	$-5.0$	$+6.3$
X-VLA	44	$+70.0$	$-7.1$	$+3.9$
OpenVLA-OFT	42	$+38.8$	$+1.9$	$+7.1$
OpenVLA-OFT	43	$+40.0$	$+3.8$	$+8.9$
OpenVLA-OFT	44	$+41.2$	$+1.2$	$+7.0$

Appendix

Table 13: Camera-view sensitivity to training seed on the fixed 560-episode subset. All values are percentage-point gains over the corresponding base model on the same episodes.

Model	$\bm{G^{\mathrm{target}}}$	$\bm{G^{\mathrm{off}}}$	$\bm{G^{\mathrm{worst}}}$	$\bm{G^{\mathrm{all}}}$
OpenVLA-OFT	$+39.3$ $[+30.6,+47.4]$	$+1.4$ $[+0.2,+2.8]$	$-2.8$ $[-5.1,-0.9]$	$+7.5$ $[+5.4,+9.6]$
$\pi_{0.5}$	$+22.0$ $[+15.2,+28.8]$	$-0.3$ $[-1.7,+1.2]$	$-3.9$ $[-6.3,-1.6]$	$+3.0$ $[+1.4,+4.7]$
InternVLA-A1.5	$+2.6$ $[-4.0,+9.2]$	$-5.2$ $[-7.3,-3.5]$	$-12.4$ $[-16.1,-8.9]$	$-4.2$ $[-6.5,-2.1]$
NORA-1.5	$+28.7$ $[+22.3,+35.4]$	$-10.1$ $[-13.9,-6.7]$	$-17.8$ $[-26.6,-11.3]$	$-3.7$ $[-7.2,-0.4]$
X-VLA	$+69.4$ $[+61.3,+76.4]$	$-4.6$ $[-7.4,-2.1]$	$-10.7$ $[-15.1,-7.1]$	$+7.0$ $[+4.3,+9.6]$

Appendix

Table 14: Camera-view tuning: summary gains and pointwise

95\%

cluster-bootstrap confidence intervals. Values are percentage points.

Model	$\bm{G^{\mathrm{target}}}$	$\bm{G^{\mathrm{off}}}$	$\bm{G^{\mathrm{worst}}}$	$\bm{G^{\mathrm{all}}}$
OpenVLA-OFT	$+19.8$ $[+13.4,+26.7]$	$+0.3$ $[-1.1,+1.7]$	$-0.6$ $[-5.0,+0.3]$	$+3.4$ $[+2.0,+4.9]$
$\pi_{0.5}$	$+10.2$ $[+6.7,+14.0]$	$-1.3$ $[-2.1,-0.7]$	$-4.1$ $[-6.2,-3.0]$	$+0.3$ $[-0.4,+1.1]$
InternVLA-A1.5	$-1.9$ $[-6.4,+2.3]$	$-2.5$ $[-3.4,-1.6]$	$-6.0$ $[-8.9,-3.7]$	$-2.6$ $[-3.7,-1.5]$
NORA-1.5	$+45.3$ $[+38.6,+51.6]$	$-1.0$ $[-3.4,+1.4]$	$-4.4$ $[-9.4,-2.6]$	$+6.2$ $[+3.7,+8.6]$
X-VLA	$+32.8$ $[+27.5,+38.0]$	$-2.2$ $[-3.8,-0.7]$	$-4.2$ $[-8.1,-2.6]$	$+3.2$ $[+1.6,+4.9]$

Appendix

Table 15: Sensor-noise tuning: summary gains and pointwise

95\%

cluster-bootstrap confidence intervals. Values are percentage points.

Evaluation dimension	Camera-view tuning	Sensor-noise tuning
Camera view	$+39.3\;[+30.6,\,+47.4]$	$+0.0\;[-4.3,\,+4.4]$
Robot state	$+0.3\;[-1.4,\,+2.1]$	$-0.1\;[-1.5,\,+1.3]$
Language	$+0.1\;[-1.4,\,+1.7]$	$-0.2\;[-1.6,\,+1.2]$
Lighting	$+1.0\;[-2.0,\,+3.9]$	$-0.6\;[-4.5,\,+2.6]$
Background	$+2.0\;[-2.5,\,+6.4]$	$+1.7\;[-1.0,\,+4.4]$
Sensor noise	$+7.9\;[+4.7,\,+11.4]$	$+19.8\;[+13.4,\,+26.7]$

Appendix

Table 16: OpenVLA-OFT: paired dimension-level changes and pointwise

95\%

cluster-bootstrap confidence intervals. Values are percentage points.

Evaluation dimension	Camera-view tuning	Sensor-noise tuning
Camera view	$+22.0\;[+15.2,\,+28.8]$	$-4.1\;[-6.0,\,-2.1]$
Robot state	$-0.5\;[-3.1,\,+2.3]$	$-3.5\;[-5.9,\,-1.2]$
Language	$-1.1\;[-3.5,\,+1.5]$	$-1.1\;[-2.7,\,+0.4]$
Lighting	$+3.0\;[+0.4,\,+6.0]$	$-0.1\;[-2.1,\,+1.8]$
Background	$+1.6\;[-1.0,\,+4.4]$	$+0.9\;[-0.3,\,+2.3]$
Sensor noise	$-3.9\;[-6.3,\,-1.4]$	$+10.2\;[+6.7,\,+14.0]$

Appendix

Table 17:

\pi_{0.5}

: paired dimension-level changes and pointwise

95\%

cluster-bootstrap confidence intervals. Values are percentage points.

Evaluation dimension	Camera-view tuning	Sensor-noise tuning
Camera view	$+2.6\;[-4.0,\,+9.2]$	$-6.0\;[-8.9,\,-3.2]$
Robot state	$-12.4\;[-16.1,\,-8.9]$	$-3.4\;[-5.3,\,-1.5]$
Language	$-5.8\;[-9.1,\,-2.8]$	$-2.8\;[-4.9,\,-0.8]$
Lighting	$-1.4\;[-3.8,\,+0.9]$	$+0.7\;[-2.5,\,+5.2]$
Background	$-3.4\;[-7.5,\,-0.3]$	$-3.0\;[-5.9,\,-0.6]$
Sensor noise	$-6.1\;[-11.6,\,-1.8]$	$-1.9\;[-6.4,\,+2.3]$

Appendix

Table 18: InternVLA-A1.5: paired dimension-level changes and pointwise

95\%

cluster-bootstrap confidence intervals. Values are percentage points.

Evaluation dimension	Camera-view tuning	Sensor-noise tuning
Camera view	$+28.7\;[+22.3,\,+35.4]$	$+0.9\;[-0.4,\,+2.2]$
Robot state	$-6.6\;[-11.4,\,-2.0]$	$-4.4\;[-6.9,\,-1.9]$
Language	$-11.9\;[-18.6,\,-5.4]$	$-3.4\;[-6.6,\,-0.3]$
Lighting	$-7.5\;[-18.2,\,+3.3]$	$+5.3\;[-2.3,\,+13.4]$
Background	$-17.8\;[-26.6,\,-9.6]$	$-2.9\;[-9.4,\,+3.3]$
Sensor noise	$-5.9\;[-7.2,\,-4.6]$	$+45.3\;[+38.6,\,+51.6]$

Appendix

Table 19: NORA-1.5: paired dimension-level changes and pointwise

95\%

cluster-bootstrap confidence intervals. Values are percentage points.

Evaluation dimension	Camera-view tuning	Sensor-noise tuning
Camera view	$+69.4\;[+61.3,\,+76.4]$	$-2.8\;[-5.1,\,-0.7]$
Robot state	$+0.5\;[-3.2,\,+4.1]$	$-3.3\;[-6.1,\,-0.5]$
Language	$-6.6\;[-10.1,\,-3.2]$	$-4.2\;[-7.9,\,-0.8]$
Lighting	$-2.2\;[-8.2,\,+3.7]$	$+1.2\;[-1.7,\,+5.9]$
Background	$-3.9\;[-9.0,\,+1.0]$	$-2.2\;[-6.5,\,+1.7]$
Sensor noise	$-10.7\;[-15.1,\,-6.7]$	$+32.8\;[+27.5,\,+38.0]$

Appendix

Table 20: X-VLA: paired dimension-level changes and pointwise

95\%

cluster-bootstrap confidence intervals. Values are percentage points.

Figure 9: Paired X-VLA trajectories under a Sensor-noise perturbation. Six ordered frames are shown for the base failure (top) and Sensor-noise-tuned success (bottom) on the same episode. Frame labels give simulator steps.

Figure 10: Paired X-VLA trajectories under a Camera-view perturbation. Six ordered frames are shown for the base failure (top) and Camera-view-tuned success (bottom) on the same episode. Frame labels give simulator steps.

VLA Model Adaptation

See all 38 papers

Not All Layers Need Tuning: Diagnosing and Directing Adaptation in Vision-Language-Action Models
Sep 16, 2026Shahram Najam Syed, Arthur Jakobsson, Prayuj Sachdev +1VLA Model Adaptation Low-Rank Adaptation
Robotics Institute, Carnegie Mellon University, Pittsburgh, USA · Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, USA
Does Adversarial Training Improve Generalization in Multi-View VLAs? Revealing and Mitigating View Collapse
Sep 27, 2026Futa Waseda, Shuhei Kurita, Isao EchizenAdversarial Training Multi-View Learning
National Institute of Informatics Tokyo, Japan
PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models
May 11, 2026Xinyu Guo, Bin Xie, Wei Chai +4VLA Model Adaptation Robot Policy Adaptation
Institute of Automation, Chinese Academy of Sciences · Dexmal · Nanjing University of Aeronautics and Astronautics +1

Sep 16, 2026J/K move · Enter open · S save

Not All Layers Need Tuning: Diagnosing and Directing Adaptation in Vision-Language-Action Models

Shahram Najam Syed, Arthur Jakobsson, Prayuj Sachdev, Jeffrey Ichnowski

Robotics Institute, Carnegie Mellon University, Pittsburgh, USA · Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, USA

VLA Model Adaptation Low-Rank Adaptation Vision-Language-Action Models Parameter-Efficient Fine-Tuning

Open paper AskPDF arXiv

Fine-tuning a Vision-Language-Action (VLA) model for a new deployment environment is expensive, yet most methods apply uniform-capacity adapters to every network region as if every region requires equal adjustment. This paper tests that assumption on five architecturally diverse VLAs (OpenVLA-OFT,

π_0

, SmolVLA, DTP, Octo; 93M-7B parameters). Measuring per-region adaptation cost as normalized parameter displacement under region-isolated fine-tuning reveals an adaptation spectrum in which appearance shifts concentrate cost in the vision encoder, instruction shifts in the language backbone, and novel-object shifts in the vision encoder together with the action head, across all five architectures. To exploit this structure, we introduce a pipeline that observes, diagnoses, allocates, and adapts. From ten unlabeled target observations and without fine-tuning, the diagnostic estimates per-region cost by combining reference-free gradient and Monte Carlo Dropout signals with a Centered Kernel Alignment score against a cached source reference; the allocator converts the estimates into variable-rank LoRA adapters under a parameter budget and freezes well-calibrated regions; and standard LoRA fine-tuning trains the resulting adapters. The diagnostic ranks regions within each deployment at a median Spearman of 0.91, and the allocation matches or exceeds uniform LoRA at every budget we tested on LIBERO and CALVIN. On a physical xArm-7, the pipeline matches full fine-tuning under an instruction-wording shift with 0.04% of its trainable parameters, and on five held-out scenes evaluated without retraining it leads every baseline, with 11-23 successes of 30 rollouts against 8-18 for the strongest parameter-efficient baseline at equal or larger budgets and 2-11 for full fine-tuning. These results suggest that adaptation cost in VLAs is structured enough to measure before fine-tuning begins.

Abstract

VLA Model Adaptation

Explore similar work

Not All Layers Need Tuning: Diagnosing and Directing Adaptation in Vision-Language-Action Models

Does Adversarial Training Improve Generalization in Multi-View VLAs? Revealing and Mitigating View Collapse

PriorVLA: Prior-Preserving Adaptation for Vision-Language-Action Models